ABSTRACT Artificial intelligence and machine learning (ML) now shape decisions in healthcare, finance and security, but they can reproduce historical prejudice and inequality. Bias in training data and in model implementation can amplify harm, especially for racial and gender minorities. Despite sustained research on fairness, mitigation in real‐world systems remains uneven, in part because stakeholders lack a shared and precise grasp of core notions, including bias, prejudice, discrimination and fairness. As a result, technical interventions are sometimes adopted without consistent conceptual grounding and reporting. This article addresses that problem by providing a knowledge base that aligns key concepts with empirical evidence and lifecycle stages. We conduct a scoping review to map sources of bias across the ML lifecycle and to identify forms of prejudice and discrimination associated with the use of sensitive attributes. We synthesize qualitative and quantitative evidence and introduce a conceptual model for organizing these findings. Our contributions are threefold: a refined lifecycle taxonomy of bias sources that introduces two additional types and spans all development stages; the explicit treatment of cognitive bias as a cross‐cutting meta‐bias; and an analysis of prejudice and discrimination that compiles a legally grounded catalogue of sensitive attributes and discusses their concepts and issues. Together, these results provide an integrated view of where and how bias emerges, and they support future research, evaluation and governance work on fairness in ML.

ABSTRACT Introduction As aging populations grow and caregiving systems strain, social robots are increasingly proposed to assist in older‐adult care. Yet, acceptance among caregivers, particularly in underrepresented regions such as Brazil, remains poorly understood. Objectives This exploratory study investigates how Brazilian caregivers perceive the utility, advantages, disadvantages, and overall acceptance of social robots in older‐adult care, accounting for demographic background, workload conditions, and prior familiarity. Method An exploratory cross‐sectional survey was administered to 94 Brazilian caregivers (formal and informal) after exposure to a realistic video demonstration of the Robios social robot. The questionnaire assessed preferences, attitudes, and acceptance across 10 structured items and 3 multiple‐choice blocks. Exploratory statistical analyses used Spearman correlations, Mann–Whitney U tests, chi‐square tests, and the Jonckheere–Terpstra trend test, with Holm‐adjusted p values for multiple comparisons. Results Caregivers prioritised functionalities related to safety monitoring and medication reminders, while expressing concerns about maintenance, technical failures, and online security. Perceived workload, rather than objective hours or caregiving experience, was positively associated with two acceptance indicators. Prior awareness of social robots was not clearly associated with acceptance in this sample. No significant differences emerged between formal and informal caregivers. Evaluation Unlike prior studies concentrated in high‐income, more institutionalised contexts, this research provides an initial systematic analysis of caregiver acceptance in Brazil's middle‐income, familistic, and predominantly informal care ecosystem. The findings challenge assumptions about demographic predictors and emphasise situational, perceptual, and cultural drivers, highlighting a collaborative caregiver‐assistant view of social robots in this context. Conclusion In this Brazilian setting, acceptance of social robots appears to be shaped less by background variables and more by perceived usefulness and relief under demanding conditions. These exploratory results inform user‐centred design and policy strategies for deploying socially adaptive robotic systems in eldercare and motivate future confirmatory studies in diverse cultural settings.

Question Answering systems based on large language models are widely employed today, benefiting from continuous enhancements and improved performance. The legal domain has become a particularly active focus for Question Answering systems, given its complexity and social importance. This paper offers a discussion on how larger and smaller language models can be used to build a legal document-based Question Answering system. We present a novel model, named Cocoruta, generated by fine-tuning with a corpus of legal documents. In addition, we examine five LLMs as they answer questions related to the legal aspects of a specific domain – the Blue Amazon, a region of particular interest involving environmental issues. The results suggest that while LLMs are not yet of sufficient quality for use as core in legal context Question Answering systems, fine-tuning on specialized corpora imparts a beneficial bias to their legal discourse. Despite having fewer parameters, the Cocoruta model competes well with larger LLMs in this aspect.

Process mining encompasses a series of tasks aimed at automatically unveiling knowledge about business processes from event logs registered in underlying information systems deployed in organizations. As well as numerous machine learning approaches, process mining approaches often require a vector space as input. However, the choice of the representational scheme to map event log information to a vector space sig-nificantly influences the quality of the results. This mapping poses challenges due to the diverse information in event logs and the intricate relationships within a business process. Relying solely on automated approaches may overlook relevant information, necessitating the incorporation of domain knowledge from external sources. Unfortunately, this incorporation introduces complexity. To address these inherent issues in constructing adequate vector spaces for process mining, this paper proposes a novel approach leveraging graph embedding to organize process-related information. To this end, we present a novel and highly flexible graph structure to represent process-related information that is then mapped to a dense vector space by applying the metapath2vec algorithm. The resulting dense vector space was compared to traditional vector spaces in an exploratory study, in which we solved the trace clustering task. We employ the N3 measure to assess the quality of the clusters and to verify whether domain knowledge is adequately represented in the vector spaces. The results demonstrate a superior potential of the dense vector spaces obtained via graph embedding to adequately organize the information to be submitted to the trace clustering task.

Process mining encompasses a series of tasks aimed at discovering knowledge about business processes from event logs underlying information systems deployed in organizations. Considering real-world business processes, high-complexity issues often prevent process mining techniques from producing satisfactory results. Business processes’ complexity arises from: (i) high behavioral variability as presented in unstructured processes, e.g. knowledge intensive processes, in which decisions commonly dependent on human actions; (ii) data volume, as it can reach big data levels in organizations with high-volume operations. Trace clustering can support mitigating high-complexity related issues. The process instance profiles resulting from trace clustering divide a complex problem into smaller and simpler ones. However, interpreting clustering results frequently requires decision-making and reasoning that might benefit from domain experts’ knowledge. Especially, in trace clustering-based process mining tasks, domain experts involvement enable results evaluation from the business process perspective. In this paper, a proposal for a trace clustering results visualization is presented. This visualization strategy supports evaluation from a business process perspective, enabling human-in-the-loop strategies. In order to illustrate the usefulness and appropriateness of the visualization, we present three use cases modeled on real-world event logs.

Current research in natural language processing is highly dependent on carefully produced corpora. Most existing resources focus on English; some resources focus on languages such as Chinese and French; few resources deal with more than one language. This paper presents the Pirá dataset, a large set of questions and answers about the ocean and the Brazilian coast both in Portuguese and English. Pirá is, to the best of our knowledge, the first QA dataset with supporting texts in Portuguese, and, perhaps more importantly, the first bilingual QA dataset that includes this language. The Pirá dataset consists of 2261 properly curated question/answer (QA) sets in both languages. The QA sets were manually created based on two corpora: abstracts related to the Brazilian coast and excerpts of United Nation reports about the ocean. The QA sets were validated in a peer-review process with the dataset contributors. We discuss some of the advantages as well as limitations of Pirá, as this new resource can support a set of tasks in NLP such as question-answering, information retrieval, and machine translation.

Although process model discovery has been extensively investigated over the past two decades, existing discovery methods are still not considered fully satisfactory. One problem is the difficulty of discovering accurate process models, achievable with both high recall (or fitness) and high precision, particularly for real-world event logs. This paper introduces a process discovery method, namely X-Processes, based on genetic algorithms, which aims to optimize accuracy through the F-Score calculated between recall and precision. Although genetic algorithms have been used to discover process models, such methods also have limitations as do other non-genetic algorithms-based methods. Experimental results for 12 real-world event logs show the accuracy of the process models discovered by X-Processes is higher than those of six other state-of-the-art discovery methods, including one also based on genetic algorithms. Besides accuracy, X-Processes delivers sound process models. Although its execution time is longer than the other compared discovery methods, X-Processes emerges as a solution when the need for a highly accurate process model outweighs the hunger for agility.

Improving judicial performance has become increasingly relevant to guarantee access to justice for all, worldwide. In this context, technology-enabled tools to support lawsuit processing emerge as powerful allies to enhance the justice efficiency. Using electronic lawsuit management systems within the courts of justice is a widespread practice, which also leverages production of big data within judicial operation. Some jurimetrics techniques have arisen to evaluate efficiency based on statistical analysis and data mining of data produced by judicial information systems. In this sense, the process mining area offers an innovative approach to analyze judicial data from a process-oriented perspective. This paper presents the application of process mining in a event log derived from a dataset containing business lawsuits from the Court of Justice of the State of Sao Paulo, Brazil - the largest court in the world - in order to analyze judicial performance. Although the results show these lawsuits have an ad hoc sequence flow, process mining analysis have allowed to identify most frequent activities and process bottlenecks, providing insights into the root causes of inefficiencies.

In recent years, machine learning has been used for data processing and analysis, providing insights to businesses and policymakers. Deep learning technology is promising to further revolutionize this processing leading to better and more accurate results. Current trends in information and communication technology are accelerating widespread use of web services in supporting a service-oriented architecture (SOA) consisting of services, their compositions, interactions, and management. Deep learning approaches can be applied to support the development of SOA-based solutions, leveraging the vast amount of data on web services currently available. On the other hand, SOA has mechanisms that can support the development of distributed, flexible, and reusable infrastructures for the use of deep learning. This paper presents a literature survey and discusses how SOA can be enabled by as well as facilitate the use of deep learning approaches in different types of environments for different levels of users.

Coclustering algorithms are an alternative to classic one-sided clustering algorithms. Because of its ability to simultaneously cluster rows and columns of a dyadic data matrix, coclustering offers a higher value-added information: it offers column clusters besides row clusters, and the relationship between them in terms of coclusters. Different structures of coclusters are possible, and those that overlap in terms of rows or columns still represent an open question with room for improvements. In addition, while most related literature cites coclustering as a means of producing better results from one-side clustering, few initiatives study it as a tool capable of providing higher quality descriptive information about this clustering. In this paper, we present a new coclustering algorithm - OvNMTF, based on triple matrix factorization, which properly handle overlapped coclusters, by adding degrees of freedom for matrix factorization that enable the discovery of specialized column clusters for each row cluster. As a proof of concept, we modeled text analysis as a coclustering problem with column overlaps, assuming that given words (data matrix columns) are associated with over one document cluster (row cluster) because they can assume different semantic relationships in each association. Experiments on synthetic data sets show the OvNMTF algorithm reasonableness; experiments on real-world text data show its power for extracting high quality information.

In this paper, we present the results obtained in a systematic study related to dimensionality reduction on gesture windowed data. The aim of the study was to analyze the effects that such reduction causes on the performance of classification models used to implement the gesture phase segmentation task. Piecewise Aggregate Approximation was used to implement the dimensionality reduction and k-Nearest Neighbors was used to implement the classification models. The results showed that the dimensionality reduction can improve the classification models' performance both by decreasing the complexity of their decision space and by improving the quality of the data.

Smart toys raises new concerns for parents and researchers. Children are more likely to share sensitive data and are unaware or rarely care about online risks. Parents play a relevant role in protecting the children, and parental control tools are necessary to take control and properly manage their child's data, according to their preferences. However, current tools neither meet parental needs nor are compliant with a standard for toy makers. We present a study of requirements for the development of a parental control tool for smart toys.

Business process descriptions are useful documents that are becoming increasingly important for identifying and documenting business processes. They are particularly beneficial during discovery when information about the process is gathered in interviews or by observation. Such business process descriptions are written as natural language text, which makes them intrinsically ambiguous. For this reason, it is the major challenge to formulate them in a precise and correct way right from the start. Therefore, this paper presents a service oriented architecture that analyzes a process description written in natural language to generate a sound process description. Being sound means that a description is structured, unambiguous, reveals possible quality and soundness problems related to BPMN 2.0, and contains clear identifiers for all known process elements in the original text. More specifically, we develop specific analysis and transformation techniques that are integrated by our proposed architecture. For validation purposes, we have implemented a prototype of this architecture. Our evaluation demonstrates that our techniques to generate sound process descriptions cover an average of 95% of the information extracted from its original process description while maintaining quality properties. Finally, our architecture can be enhanced with additional services that contribute to the creation and management of processes descriptions in organizations.

Process model discovery can be approached as an optimization problem, for which genetic algorithms have been used previously. However, the fitness functions used, which consider full log traces, have not been found adequate to discover unstructured processes. We propose a solution based on a local analysis of activity transitions, which proves effective for unstructured processes, most common in organizations. Our solution considers completeness and accuracy calculation for the fitness function.

Few approaches allow assertive estimates for ticket completion time in incident management.The accuracy level of prediction models depends on how useful the used attributes are.Moreover, to effectively use computational resources, a canonical attribute subset must be used.This paper proposes two automated attribute selection methods to build prediction model.A filter method and two wrapper search techniques were combined with annotated transition systems to automate attribute selectors applied to a real-life incident management process.The results show that the wrapper method surpassed human experts' decision making.

The increasing number of moving vehicles along roads and the lack of supporting infrastructure is a wellestablished problem. Major consequences are augmenting of traffic jams, accidents, fuel consumption and pollution. Vehicular Ad hoc NETworks (VANETs) represent opportunities to deal with the aforementioned problems. In VANETS, efficiency and safety to applications are provided using communication support. In efficiency applications, each vehicle is aware of its location. Using this information and communication support, vehicles collaborate to reduce travel time and to improve mobility. In contrast, safety applications aim to reduce or even avoid accidents, and must obey strong timing constraints. In this context, VANETs applications can benefit from Computational Intelligence (CI) and adaptive approaches to implement the required demands. Thus, the contribution of this paper is twofold: $( i)$ we discuss how VANETs can benefit from CI and Artificial Intelligence techniques to make transportation networks more efficient regarding to safety applications, and, $( ii)$ we report our current work and new directions in the development of efficiency applications to VANETs using adaptation and CI techniques.

Modern techniques widely applied in data mining, including computational intelligence and machine learning, have been fairly neglected in process mining. We conducted an exploratory study to use artificial neural networks to extract knowledge from an unstructured process in the distance learning domain. We discuss some possible benefits and limitations regarding the mining of unstructured processes. Results suggest that applying either classical process mining or modern data mining techniques would result in significant benefits for this domain. Our work helps to guide new studies related to the application of modern mining techniques in process mining.

One reason for researching new biometric modalities is to improve the capabilities of security systems against threats. Biometric modalities based on biomedical signals, in particular the electrocardiogram signal (ECG), have been widely adopted. These can be represented by time series. However, in this context, a critical issue is how to extract features from ECG signals effectively. Several techniques have been put forward regarding the best way to represent time series, in particular techniques that are based on symbolic values, and significant results have been achieved by means of these techniques for addressing different types of problems. In this paper, we present twenty symbolic representations of time series applied to the extraction of nonfiducial features from ECG signals that aim at biometric recognition. In addition, we put forward three novel symbolic representations, namely, Representation based on Kmeans (R-Kmeans), Symbolic Aggregate approXimation based on Kmeans (SAX-Kmeans), and Extended Symbolic Aggregate approXimation based on Kmeans (ESAX-Kmeans). Experimental results conducted on two publicly available datasets indicate that the novel representations can improve the performance of recognition compared with the others.

Currently the demand for the development of more precise and reliable methods of person identification have received attention from the academic community and industry, with Biometrics being one of these new approaches. The term ‘Biometrics’ is used to refer to identification techniques based on physical or behavioural characteristics. As biometric recognition becomes increasingly popular, the fear of circumvention, obfuscation and replay attacks is a rising concern. Since the traditional biometric modalities (face, iris and fingerprint) are not able to supply the needs of every possible security requirement, numerous emerging biometric modalities are presented, trying to fill the gap. Biomedical signals, like electrocardiogram (ECG) and electroencephalogram (EEG), have been proposed as emerging biometric modalities. The advantages of using the ECG for biometric recognition can be summarized as universality, permanence, uniqueness, robustness to attacks, liveness detection. According to the utilized features, the existing ECG based biometric systems can be classified to fiducial, non-fiducial and hybrids systems. This papers analyses the impact of some feature selection strategies like Genetic Algorithm, Memetic Algorithm and Particle Swarm Optimization on the performance of Biometric Systems based on ECG using K-Nearest Neighbours, Support Vector Machines, Optimum Path Forest and a Euclidean Distance Classifier for classification task. The results show that there is a subset of features extracted from the ECG signal that provides high recognition rates.

In recent years, human identification based on face recognition has attracted the attention of the scientific community and the general public due to its wide range of applications. A face recognition system involves three important phases: face detection, feature extraction and classification (identification and/or verification). The robustness of face recognition could be improved by treating the variations in these stages. One of the main issues in design of face recognition system is how to extract discriminative facial features. A precise extraction of a representative feature set will improve the performance of a face recognition system. Various techniques have been used to represent images efficiently, of which the most well-known and widely applied are Wavelet, Contourlet, Shearlet and Curvelet Transform. Their ability to capture localized time-frequency information of image motivates their use for feature extraction. In this paper, we conduct a systematic empirical study on these transforms as feature extractors from face images. To further reduce the feature dimensionality, we adopt Principal Component Analysis and Linear Discriminant Analysis to select the most discriminative feature sets. The performance levels delivered by each transform are contrasted in terms of the accuracy measure computed over the outputs generated by the Support Vector Machine classifier (SVM). Experimental results conducted on a publicly available database are reported whereby we observe that the Curvelet Transform followed by the Wavelet Transform significantly outperform the others according to accuracy measure calculated over the SVM classifier.

Camera calibration is an attractive problem that have received relatively great attention from scientists in the last years. It has several interesting applications, particularly in sports broadcasting technology. We present in this paper a complete system that receives a sport image, automatically detects control points in the image with a RANSAC-based strategy and solves the optimization problem by using a genetic algorithm. In fact, our approach introduces an improvement in the RANSAC-based strategy which is able to refine its results, and shows that a simple genetic algorithm is capable to solve the camera calibration problem with relatively small computational effort.

This paper describes a hybrid architecture that provides automatic classification for a set of gestures. Such architecture combines fuzzy-connectionist, heuristic and syntactical pattern recognition approaches, and deals with gesture recognition based on primitives. The modeling with primitives allows the use of multiples classifiers in order to achieve high classification accuracy. The heuristic classifier and the fuzzy syntactical integrating strategy are described in this paper. The fuzzy-connectionist classifiers were discussed in previous works and they are now revisited just to present the set of parameters that solves the current proof of concept, in the scope of Brazilian Sign Language Manual Alphabet. The fuzzy syntactical strategy coupled with the modeling with primitives has improved the pattern recognition results, enabling the design of architecture for classification with high flexibility and scalability to development of applications in different signed communication contexts. The experimental results show that the proposed approach is valid and has promising application.

Recently, Support Vector Machines have presented promissing results to various machine learning tasks, such as classification and regression. These good results have motivated its application to several complex problems, including temporal information analysis. In this context, some studies attempt to extract temporal features from data and submit these features in a vector representation to traditional Support Vector Machines. However, Support Vector Machines and its traditional variations do not consider temporal dependency among data. Thus, some approaches adapt Support Vector Machines internal mechanism in order to integrate some processing of temporal characteristics, attempting to make them able to interpret the temporal information inherent on data. This paper presents a review on studies covering this last approach for dealing with temporal information: incorporating temporal reasoning into Support Vector Machines and its variations.

This paper presents an approach for carrying out gesture recognition for the Brazilian Sign Language Manual Alphabet. The gestural patterns are treated as a combination of three primitives, or cheremes - hand configuration, hand orientation and hand movement. The recognizer is built in a modular architecture composed by inductive reasoning modules, which use the artificial neural network Fuzzy Learning Vector Quantization; and rule-based modules. This architecture has been tested and results are presented here. Some strengths of such approach are: robustness of recognition, portability to similar contexts, extensibility of the dataset to be recognize and reduction of the vocabulary recognition problem to the recognition of its primitives.

In this paper, the vision-based hand movement recognition problem is formulated for the universe of discourse of the Brazilian Sign Language. In order to analyze this specific domain we have used the artificial neural networks models based on distance, including neural-fuzzy models. The experiments explored here show the usefulness of these models to extract helpful knowledge about the classes of movements and to support the project of adaptative recognizer modules for Libras-oriented computational tools. Using artificial neural networks architectures - Self Organizing Maps and (Fuzzy) Learning Vector Quantization, it was possible to understand the data space and to build models able to recognize hand movements performed for one or more than one specific Libras users.

Self organizing map (SOM) is a kind of artificial neural network with a competitive and unsupervised learning. This technique is commonly used to dataset clustering tasks and can be useful in patterns recognition problems. This paper presents an artificial neural network application to signals language recognition problem, where the image representation is given by bit signatures. The recognition results are promising and are presented in this paper. More, some analysis about the combination ldquoSOM + bit signaturerdquo improved our understanding about the characteristics of the LIBRAS signals and the conclusions are also listed in this paper.

This paper presents the principal results of a detailed study about the use of the Meaningful Fractal Fuzzy Dimension measure in the problem in determining adequately the topological dimension of output space of a Self-Organizing Map. This fractal measure is conceived by combining the Fractals Theory and Fuzzy Approximate Reasoning. In this work this measure was applied on the dataset in order to obtain a priori knowledge, which is used to support the decision making about the SOM output space design. Several maps were designed with this approach and their evaluations are discussed here.